Detection And Automatic Classification Of Appearance Defects Of Polarizers Based On Dark Field And Telecentric Imaging

As a common polarizing optical element,the polarizer is one of the core components of the display panel and has extremely wide applications.The quality of the polarizer has an important impact on the yield of the display panel.If the polarizer with appearance defects flows into the assembly process,the entire panel may be scrapped.Therefore,regardless of whether it is a polarizer manufacturer or a display panel manufacturer,strict inspection of polarizer appearance defects is required.In the production process,after the defect is detected,the quality personnel must also analyze the cause of the defect to improve the production process and process in time.Therefore,it is necessary to classify the detected defect samples.In the current production,offline manual visual inspection is mainly used to detect the appearance defects of polarizers,and then these defective samples are manually marked and classified.The method of manual inspection is affected by the subjective judgment of the quality inspection personnel.There are problems such as missed inspection,false inspection,and low inspection efficiency,which are difficult to meet the needs of modern production.Therefore,the development of a new automatic detection and classification technology for the appearance defects of polarizers in order to further improve the detection accuracy and detection speed is an urgent problem to be solved in the field of polarizer production.The main contents of this paper are as follows:1.Researched a method for detecting the appearance defects of polarizers based on dark field imaging.Among them,the detection device contains two main modules of lighting and image acquisition.Through the dark field imaging method,the defect of the polarized film test sample is "bright" after imaging,and the background is "dark",which can effectively enhance the contrast of the defect and thereby improve the defect detection rate.For the key influencing factor in the detection method-the incident angle of light,a comparative experimental test was conducted in this paper.From the relevant experimental results,the optimal incident angle of light was determined to be 15?~ 30?.Using the detection methoddescribed,this article tested 5 common types of polarizer appearance defects samples,collected 200 pieces of each type,a total of 1000 dark field polarizer appearance defect images,and composed of dark field polarizer appearance defect basic data set.2.A method for detecting the appearance defects of polarizers based on telecentric imaging is studied.In view of the shortcomings of the existing polarizer appearance defect detection methods,this article uses the telecentric imaging method to detect the polarizer appearance defects,which can obtain richer defect details,provide more basis for the judgment of the cause of the defect,and provide manufacturers a new polarizer quality analysis method,so as to improve the production process in time,optimize the production process,and improve the product yield.Finally,by the detection method collected 12 categories,30 images of each category,a total of 360 telecentric polarizer appearance defect images,which constituted a telecentric small sample polarizer appearance defect data set.3.The data enhancement method based on generative adversarial network is studied,and an improved generative adversarial network: LOG-Pix2 Pix is proposed.In order to solve the problem that the number of dark-field polarizer appearance defect images is insufficient to effectively train the deep learning network for classification,this paper uses the method of generating defect images based on the generated confrontation network to expand the original data set.Aiming at the blur of the generated image in the Pix2 Pix model,LOG variance is introduced to improve its objective function,and the resulting LOG-Pix2 Pix model can generate higher quality images.In this paper,1000 images of each category are generated through the LOG-Pix2 Pix model,a total of 5000 "false" dark field polarizer appearance defect images,and then added to the dark field polarizer appearance defect basic data set to form dark field polarizer appearance defect mixed data set,and then train and classify the deep learning network.4.An improved fusion classification network based on Res Net18 is proposed.Based on the Res Net18 classification network,this paper first introduces deformable convolution to improve it,so that it can better perform convolution operation on the entire defect area to extract more effective feature information for classification.Then,the ordinary 3 * 3convolution in the model is replaced with the ACB convolution module.In the case of equivalent calculations,the classification accuracy of the model can be further improved.Finally,get the final fusion classification network:Res Net18-DACB.The classification results on the dark field polarizer appearance defect hybrid dataset and the three public datasets Cifar10,Cifar100,and Mini-Imagenet show that the fusion classification network can effectively improve the classification accuracy.5.In view of the small sample classification problem of the appearance defect images of telecentric polarizers,this paper studies three mainstream deep learning methods: deep transfer learning,meta-learning,and meta-transfer learning.Through actual model training and testing,the advantages and disadvantages of the three methods are compared and analyzed.In the study of unsupervised classification based on meta-transfer learning,in order to allow the meta-classifier to better learn how to classify the appearance defects of telecentric polarizers,this paper collected different types of defect images from some public defect data sets and constructed the MDIML,a data set for transfer learning defect classification.The experimental results show that after adding MDIML images for training,the classification accuracy of the meta-transfer classification model for the appearance defects of telecentric polarizers has been improved.